JP2009116400A - Information processing apparatus and information processing method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To flexibly change a combination between extraction processing speed and extraction accuracy when extracting a specific object from image data by use of a plurality of weak discriminators connected in cascade. <P>SOLUTION: An information processor for processing the data by use of the weak discriminators connected in cascade includes: a ROM (Read Only Memory) 205 storing information for prescribing processing contents of each of the plurality of weak discriminators; a means 201 selecting the weak discriminator to be used for the processing of the data from the plurality of weak discriminators by referring to a table prescribed with information for determining the weak discriminator to be used among the plurality of weak discriminators; and a means 203 extracting the object from the data by use of an evaluation value of the weak discriminator obtained by processing the data based on the information for prescribing the processing contents corresponding to the selected weak discriminator. <P>COPYRIGHT: (C)2009,JPO&INPIT

Description

  The present invention relates to an information processing technique for processing data using a plurality of weak classifiers connected in cascade.

  In recent years, various methods for extracting a specific object such as a human face from image data have been proposed and put into practical use.

  Among them, the technique disclosed in Non-Patent Document 1 has attracted attention because of its high speed. This method connects a discriminator composed of a plurality of weak discriminator groups generated by a boosting learning algorithm (details are disclosed in Non-Patent Document 2) in a cascade, and performs an abort determination for each discriminator. It is a method of proceeding while processing.

  FIG. 24 shows a schematic configuration of the technique disclosed in Non-Patent Document 1. 2401 to 240n (n is a natural number) are discriminators (also referred to as stages) generated by learning, and each discriminator is composed of, for example, a plurality of rectangular filters having a low processing load. Each rectangular filter used at this time is generally called a “weak classifier” because it has a low discrimination capability.

  FIG. 25 is a diagram illustrating an example of a rectangular filter for extracting a specific object. 2501a to 2501c are examples of image blocks to be extracted by a rectangular filter, and are partial images of a predetermined size cut out from the entire image data.

  FIG. 26 shows image data to be processed. Of these, 2601 represents one frame of image data to be processed. Reference numeral 2602 denotes a processing block which is a processing unit when processing one frame of image data to be processed, and is a partial image having a size corresponding to the size of the rectangular filters 2502a to 2501c. Using the rectangular filters 2502a to 2502c, the feature of the local area in the partial image is extracted by calculating the difference between the total area data indicated by the white area and the black area.

  The entire image data 2601 is sequentially scanned by the partial image 2602 with a predetermined step size, and a specific object in the image data 2601 is extracted by performing processing in the processing unit.

  Each of the discriminators 2401 to 240n accumulates evaluation values, which are output of determination results for each rectangular filter, and determines whether or not a specific object exists by performing threshold processing using an identification threshold. As described above, the discriminators 2401 to 240n are connected in cascade, and each discriminator proceeds to processing by the subsequent discriminator only when it is determined that the specific object exists in the partial image 2602 of the processing unit.

This method determines whether the partial image of the processing unit is a specific object or a non-specific object for each discriminator called a stage in this way, and if it is determined as a non-specific object, it immediately calculates. It is characterized by ending. In actual image data, since it is often determined that the object is a non-specific object at an early stage of the stage, high-speed extraction processing can be realized.
Viola, P.M. Jones, M .; “Rapid object detection using a boosted cascade of simple features”. Proc. of Computer Vision and Pattern Recognition, 2001-12, IEEE Computer Society. p. 511-518. Yoav Freund and Robert E. Shapire. A decision-theoretic generation of on-line learning and an application to boosting. In Computational Learning Theory: Eurocolt '95, pages 23-37. Springer-Verlag, 1995. JP 2005-100121 A

  Here, consider a case where the function of extracting a specific object is realized in an embedded device or the like using a technique such as that described in Non-Patent Document 1. In such a case, there are cases where it is necessary to adjust the trade-off between the extraction accuracy and the extraction processing speed in accordance with the extraction purpose and the calculation performance of the embedded device.

  As an example, a case where a hardware unit (or integrated circuit) as a common device is mounted on a plurality of embedded devices having different specifications can be cited. In such a case, it is desirable to adjust the trade-off between the extraction accuracy and the extraction processing speed in accordance with the operation clock, usage conditions, and the like of the embedded device to be mounted.

  Further, even when mounting on the same embedded device, the required extraction accuracy and extraction processing time may differ depending on the type of application mounted on the embedded device. Even in such a case, if the trade-off between the extraction accuracy and the extraction processing speed can be adjusted, the performance of the embedded device can be optimized.

  On the other hand, conventionally, for such cases, the resolution of image data to be processed is lowered, or the step size when scanning image data to be processed by a partial image is coarsened. It corresponded. For example, Non-Patent Document 1 discloses a method for changing the step size.

  However, in such a case, there is a problem that a flexible trade-off between the extraction accuracy and the extraction processing speed cannot be realized. For example, in order to control the extraction processing speed by changing the resolution, the input image data must be converted into a corresponding resolution. It is also necessary to prepare a plurality of discriminators corresponding to each resolution in advance.

  Even when the step size is changed, the types of trade-off points (appropriate combination of extraction accuracy and extraction processing speed) are limited (eg, the step size is set to n pixels) and are flexible. There is a problem that control is not possible.

  On the other hand, as in Patent Document 1, learning is performed in advance using a plurality of discriminators in consideration of the amount of calculation in an embedded device to be applied, and a plurality of discriminator groups are subjected to extraction processing. It is conceivable to cope by selecting a discriminating device group corresponding to the embedded device. However, in that case, it is necessary to prepare a plurality of different discriminator groups for each trade-off point, and there is a problem that resources such as a memory increase. This problem is particularly serious when there is a large amount of parameter information that requires a large number of classifiers.

  The present invention has been made in view of the above problems, and a combination of extraction processing speed and extraction accuracy in extracting a specific object from input data using a plurality of weak classifiers connected in cascade. The purpose is to be able to change flexibly.

In order to achieve the above object, an information processing apparatus according to the present invention comprises the following arrangement. That is,
A plurality of weak classifiers connected to the cascade to process the data,
First storage means for storing information defining the processing content of each of the plurality of weak classifiers;
Preconditions for processing the data and information for determining weak classifiers to be used for processing the data among the plurality of weak classifiers are stored in a table defined for each precondition. Second storage means for
Selecting means for selecting, from the plurality of weak classifiers, weak classifiers to be used for processing the data by referring to the table;
Based on the information that defines the processing content corresponding to the weak classifier selected by the selection means, an evaluation value obtained by processing the data by the weak classifier is used. Extraction means for extracting.

  According to the present invention, when a specific object is extracted from input data using a plurality of weak classifiers connected in cascade, the combination of the extraction processing speed and the extraction accuracy can be flexibly changed. It becomes.

  Details of each embodiment will be described below with reference to the drawings.

[First Embodiment]
<Hardware configuration of pattern identification device>
FIG. 1 is a diagram illustrating a hardware configuration example of an information processing apparatus (pattern identification apparatus 100) according to the first embodiment of the present invention. The pattern identification device 100 has a function of extracting a specific object (image pattern) in image data.

  In FIG. 1, an image input unit 101 includes an optical system, a photoelectric conversion device, a driver circuit that controls the photoelectric conversion device, an AD converter, a signal processing circuit that manages various image corrections, a frame buffer, and the like. Here, the photoelectric conversion device includes a charge-coupled devices (CCD) or a complementary metal oxide semiconductor (CMOS) sensor.

  Reference numeral 102 denotes a preprocessing unit that performs various types of preprocessing for effectively performing the extraction processing. Specifically, image data conversion such as color conversion processing / contrast correction processing is processed by hardware based on the color conversion table and contrast correction information.

  An identification processing unit 103 is a hardware block that executes central processing for realizing the information processing method according to the present embodiment. Details of the identification processing unit 103 will be described later with reference to FIG.

  Reference numeral 106 denotes a DMAC (Direct Memory Access Controller). The DMAC 106 manages data transfer between processing units on the image bus 104 and between devices on the image bus 104 and a RAM 110 / ROM 109 on a CPU (Central Processing Unit) bus 107.

  A bridge 105 provides a bridge function between the image bus 104 and the CPU bus 107. Reference numeral 108 denotes a CPU that controls the overall operation of the pattern identification apparatus 100. Reference numeral 109 denotes a ROM (Read Only Memory), which stores instructions that define the operation of the CPU 108 and data necessary for various calculations.

  Reference numeral 110 denotes a memory (RAM: Random Access Memory) necessary for the operation of the CPU 108. The RAM 110 is configured by a memory having a relatively large capacity such as a DRAM (Dynamic RAM).

  The CPU 108 can access each processing unit on the image bus 104 via the bridge 105. By separating the image bus 104 and the CPU bus 107, the hardware processing units (101 to 103) and the CPU 108 can be operated simultaneously.

  An interface unit 111 is an input / output interface for the operator to specify the operation of the pattern identification apparatus 100.

<Detailed Configuration of Identification Processing Unit 103>
FIG. 2 is a diagram for explaining the details of the identification processing unit 103. A rectangular filter calculation unit 201 executes a predetermined rectangular filter calculation process on the processing target image data 210 stored in the RAM 204. A RAM 204 is a memory for storing the image data 210 to be processed, and is composed of a high-speed memory with low latency. Reference numeral 205 denotes a ROM that stores parameter information 211 that defines the calculation contents of each rectangular filter (first storage means).

  The rectangular filter calculation unit 201 accesses the image data 210 on the RAM 204 according to rectangular coordinate information (details will be described later) of the parameter information 211 stored in the ROM 205, and performs calculation processing and weak discrimination processing using a rectangular filter generated by learning in advance. Are executed sequentially.

  Reference numeral 202 denotes a processing number counter, which includes a counter for counting a rectangular filter number being processed and a stage number being processed. A processing execution determination unit 203 controls the rectangular filter arithmetic unit 201 according to the rectangular filter processing number counter and the stage processing number counter included in the processing number counter 202 and whether or not the rectangular filter specified from the outside is executed and the end condition. To do.

  In other words, a rectangular filter designated by the processing execution determination unit 203 is used for actual processing by the rectangular filter calculation unit 201 from among a plurality of rectangular filters obtained by learning in advance.

  Reference numeral 206 denotes an internal local bus of the identification processing unit 103, which is connected to the image bus 104 via a bridge (not shown). The CPU 108 accesses various devices (201 to 205) on the internal local bus 206 via the bridge.

<Detailed Configuration of Rectangular Filter Operation Unit 201>
FIG. 3 is a diagram showing a detailed configuration of the rectangular filter calculation unit 201. FIG. 3 corresponds to a diagram for explaining the contents of 2401 to 240n in FIG. 311 to 31k are discriminators, and a plurality of rectangular filters execute arithmetic processing and weak discrimination processing according to the parameter information 211 stored in the ROM 205. Reference numeral 320 denotes a discrimination processing unit, which performs threshold processing on an accumulated value of evaluation values, which is a result of the weak discrimination processing of each of the rectangular filters 311 to 31k, using an identification threshold, and determines whether or not an object to be extracted exists.

<Configuration of Parameter Information 211>
FIG. 4 is a diagram illustrating an example of the parameter information 211 stored in the ROM 205. FIG. 4 shows an example of parameter information necessary for one stage process.

  In FIG. 4, rectangular coordinates 402 are four vertex coordinate data corresponding to one rectangular block. The coordinate data here corresponds to the coordinate value in the object to be extracted. The rectangular coordinates 402 have a set of coordinate data corresponding to the number of rectangles 401.

  The filter threshold 404 is a threshold for weakly determining the target rectangular filter output. The reliability indicates the reliability (contribution rate) to the evaluation value that is the result of the target rectangular filter. Each of the rectangular filters 311 to 31k outputs, as its output, a value obtained by multiplying the result (1 or -1) obtained by weakly discriminating the result of the rectangular filter calculation processing with the filter threshold value. The identification threshold value 405 is a threshold value for executing identification processing by the discrimination processing unit 320.

  The above parameter information 211 is stored for the number of rectangular filters (k in FIG. 4), and further stored for the number of stages.

<Processing Flow in Pattern Identification Device 100>
Hereinafter, the operation of the pattern identification apparatus 100 according to the present embodiment will be described in detail with reference to the flowchart shown in FIG. FIG. 5 describes the operation of the CPU 108 and the operation of the identification processing unit 103 on one flowchart. A flow (steps S <b> 508 to S <b> 518) specified as “identification processing” in the drawing corresponds to the operation of the identification processing unit 103.

  First, in step S501, the CPU 108 executes various initialization processes prior to the start of the identification process. Each processing unit (101 to 103) initializes an internal register, memory, and the like in accordance with an initialization instruction from the CPU.

  In step S502, a rectangular filter to be used for actual processing is specified, and the contents of the rectangular filter processing in each stage are determined. Specifically, the number of rectangular filters to be processed is designated for each stage.

  In step S503, the stage to be actually processed is designated. Specifically, the number of stages to be processed is specified. Here, the number of rectangular filters and the number of stages to be processed are determined according to the operation clock of the pattern identification apparatus 100.

  FIG. 6 is a diagram schematically showing the relationship between the total number of weak classifiers (rectangular filters) and the extraction accuracy. The total number of weak classifiers is the sum of the weak classifiers included in each stage, and erroneous detection indicates the degree of detection error.

  As shown in FIG. 6, in general, as the total number of weak classifiers is increased, false detections are reduced and the extraction accuracy is improved. On the other hand, since the total number of weak classifiers is considered to be roughly proportional to the extraction processing time, the trade-off between extraction processing speed and extraction accuracy can be adjusted very easily by controlling the total number of weak classifiers. .

FIG. 7 is an example of a table that associates the operation clock with the total number of weak classifiers to be used. In FIG. 7, N1 to N3 indicate the number of stages when the operation clock is 1 to 3, respectively, and S _xy (x: clock type, y: stage number) indicates the number of weak classifiers in each stage.

N1~N3 and S _xy, respectively, the number of stages and the weak discriminator number following values generated by learning is set. Th _xy (x: clock type, y: stage number) is an identification threshold value for each stage. The discriminator generated by the boosting learning algorithm performs identification using the accumulated value shown in Expression (1).

  Here, T is the number of weak classifiers that are actually used. When the number of weak classifiers is reduced, the number is the number after reduction. The value of T may be different for each stage.

α _i is the reliability of the weak classifier and is determined by learning of the weak classifier. α _i can also be said to be the contribution ratio of the corresponding weak classifier to the classifier.

h _i (x) is an evaluation value that is an output of the weak discriminator and is a result of weakly discriminating the outputs of the rectangular filters 2502a to 2502c shown in FIG. When the number of weak discriminators is reduced, the value of T shown in Equation (1) changes compared to that at the time of learning. Therefore, it is necessary to prepare a new discrimination threshold Th _xy for each predetermined condition. When the value of T is the same at the time of learning and at the time of execution, the identification threshold shown in FIG. 4 is used as it is.

  The table shown in FIG. 7 is created in advance by an external workstation or the like and stored in the ROM 109 or the like (second storage means). That is, the table is created so that many weak classifiers are used when the operation clock is high, and a small number of weak classifiers are used when the operation clock is low.

  Note that the combination of the reduction in the number of stages and the reduction in weak classifiers in the stage may be determined by a conventionally proposed optimization method or the like. For example, a test data set is used in advance, and a desired combination is created from the relationship between extraction accuracy and extraction processing time. By setting the number of stages and the number of weak classifiers in the stage independently, the trade-off point can be determined more flexibly.

  Furthermore, when determining the number of weak classifiers, it is possible to reduce the number of weak classifiers while suppressing a decrease in extraction accuracy by using the reliability of the weak classifiers. A weak discriminator having a low reliability has a small influence on the discrimination, and the reliability decreases as the latter stage is reached. For this reason, for example, among the weak discriminators included in the continuous discriminators, when the reliability of the weak discriminators included in the subsequent stage discriminators is a predetermined value or less, all of the plurality of discriminators are reduced, etc. The method may be used. The effect that optimization becomes easy by using the reliability is obtained.

  In steps S502 and S503, the table shown in FIG. 7 is referred to, and the number of rectangular filters, the number of stages, the identification threshold value, and the like used for actual processing are set in a register (not shown) built in the processing execution determination unit 203.

  When various settings for the identification processing are completed, the image input unit 101 is activated in step S504, and the preprocessing unit 102 is activated in step S506. Then, the following processes are executed respectively.

  First, the image input unit 101 stores image data for one frame acquired by the photoelectric conversion device in an internal buffer (not shown).

  When the storage is finished, an interrupt to the effect that the acquisition of image data is finished is generated for the CPU 108. When detecting an interrupt, the CPU 108 activates the DMAC 106 and transfers the acquired image data to an internal memory (not shown) of the preprocessing unit 102.

  The preprocessing unit 102 that has received the image data starts preprocessing. Specifically, image data conversion processing such as color conversion processing of image data and contrast correction processing is executed based on a color conversion table and contrast correction information designated in advance. Further, the preprocessing unit 102 transmits an interrupt signal to the CPU 108 when the image data conversion process is completed.

  Subsequently, in step S506, the CPU 108 detects the interrupt signal transmitted from the preprocessing unit 102, and generates Samed Area Table data (hereinafter referred to as SAT data) based on the image data after the image data conversion process. The SAT data is a data format suitable for processing a rectangular filter as shown in FIG. 25 at high speed. Note that the SAT data is disclosed in Non-Patent Document 1 and is not related to the features of the present invention, and thus detailed description thereof is omitted here.

  FIG. 8 is a diagram for explaining a method of calculating a rectangular block from SAT data. The sum of data in the rectangular area 802 in the input data area 801 is calculated by S4 + S1- (S2 + S3) using the values S1 to S4 (same coordinate positions as the input data) of the four points in the SAT data area. Can do. By referring to the SAT data, for example, in the case of the rectangular filter indicated by 2502a in FIG. 25, the rectangular filter calculation processing can be performed with 6 points of data read.

  In step S507, the CPU 108 transfers the generated SAT data and stores it in the RAM 204 of the identification processing unit 103 as image data to be processed. When the transfer of the SAT data to the RAM 204 is completed, the identification processing by the identification processing unit 103 is started.

<Flow of identification processing in identification processing unit 103>
Hereinafter, a flow of identification processing (hardware processing) of the identification processing unit 103 shown in steps S508 to S518 will be described.

  In step S508, a processing block serving as a processing unit is determined from the image data stored in the RAM 204. The processing block is a partial image indicated by reference numeral 2602 in FIG. 26, and an extraction target object included in one frame of image data is extracted by performing extraction while scanning the window in units of one pixel or one line.

  In step S509, the rectangular filter calculation unit 201 reads the parameter information 211 on the ROM 205. Furthermore, in step S510, rectangular filter calculation processing is executed with reference to predetermined image data on the processing block. As described above, the rectangular filter calculation process calculates difference values of a plurality of rectangular blocks with respect to the processing block.

  In step S511, the output of the rectangular filter in step S509 is compared with the rectangular filter threshold for the filter stored in the ROM 205. If the output of the rectangular filter exceeds the filter threshold value, the evaluation value obtained by integrating the reliability of the rectangular filter is added to the accumulated value for the rectangular filter output.

  On the other hand, when the output of the rectangular filter is equal to or less than the filter threshold, the evaluation value obtained by integrating the reliability of the rectangular filter with respect to the output of the rectangular filter is subtracted from the accumulated value.

  When one rectangular filter calculation (weak discriminator calculation) process is completed, in step S512, the processing number counter 202 updates the rectangular filter processing number counter.

  Next, in step S513, it is determined whether or not the rectangular filter calculation processing for all weak discriminator rows in one stage has been completed. Here, the processing execution determination unit 203 compares the number of rectangular filters specified in step S502 with the value of the rectangular filter processing number counter output from the processing number counter 202, and whether the rectangular filter calculation processing in the stage is completed. Determine whether or not.

  If it is determined in step S513 that the rectangular filter calculation processing using all the rectangular filters has not been completed, the process returns to step S509 to execute the rectangular filter calculation processing using the next rectangular filter.

  On the other hand, if it is determined in step S513 that the rectangular filter calculation processing for all rectangular filters has been completed, the process proceeds to step S514, where the discrimination processing unit 320 identifies the accumulated value and cancels the stage. Decide whether or not.

  Here, it is determined whether or not the extraction target object exists in the processing block by comparing with the identification threshold value 405 set in step S502. As a result of identification by the discrimination processing unit 320, when it is identified that there is no extraction target object, the identification process is terminated.

  When the identification process is terminated, in step S517, it is recorded that no extraction target object exists in the processing block (that it is a non-extraction target object).

  On the other hand, as a result of identification by the discrimination processing unit 320, when it is identified that there is an object to be extracted, the process proceeds to step S515, and the stage number counter of the processing number counter 202 is updated.

  In step S516, the process execution determination unit 203 compares the stage process number counter with the number of stages specified in step S503. As a result of the comparison in step S516, if it is determined that the specified number of stages has not been reached, the process returns to step S509 to start rectangular filter calculation processing included in the next stage.

  On the other hand, if it is determined that the specified number of stages has been reached as a result of the comparison in step S516, the process proceeds to step S517 to record that the extraction target object exists in the processing block.

  In step S518, it is determined whether identification processing has been executed for all processing blocks in the image data to be processed. If it is determined that the identification processing has not been completed for all the processing blocks in the image data to be processed, the process returns to step S508 to determine a processing block adjacent to one pixel or one line, and the same processing is repeated.

  On the other hand, if it is determined in step S518 that the identification process has been completed for all the processing blocks in the image data to be processed, the process proceeds to step S519, where the identification process is performed on the image data of the next frame. It is determined whether or not. When the identification process is executed for the image data of the next frame, the process returns to step S504, and the processes of steps S504 to S518 are repeated. On the other hand, when it is determined that the identification process is not performed on the image data of the next frame, the process ends.

  As described above, in the present embodiment, the number of weak discriminators to be used and the number of stages (corresponding to numbers indicating the order of processing) are designated according to the operation clock of the pattern identification device that performs the identification processing. It was set as the structure to do. Thereby, in the pattern identification device that executes the identification process, the rectangular filter calculation process is not executed in the stages after the designated number of stages in the identification process.

  In other words, by configuring the number of weak classifiers according to the operation clock, even when the identification processing unit is installed in multiple embedded devices with different specifications, extraction is performed according to the operation clock of the embedded device. It is possible to adjust the accuracy and the extraction processing speed.

[Second Embodiment]
In the first embodiment, the configuration for adjusting the trade-off between the extraction accuracy and the extraction processing speed by specifying the number of weak classifiers or the number of stages to be used has been described. However, the present invention is not limited to this. Absent. For example, a configuration may be adopted in which the trade-off between the extraction accuracy and the extraction processing speed is adjusted by individually specifying whether or not the weak classifier is executed based on the reliability.

  Hereinafter, the details of the pattern identification apparatus according to the present embodiment for adjusting the trade-off between the extraction accuracy and the extraction processing speed will be described. Note that the hardware configuration of the pattern identification apparatus according to the present embodiment, the detailed configuration of the rectangular filter calculation unit 201, the configuration of the parameter information 211, and the like are the same as those in the first embodiment, and thus description thereof is omitted here. .

  FIG. 9 is a flowchart showing the operation of the pattern identification apparatus 100 according to the second exemplary embodiment of the present invention. Here, only the differences from the first embodiment will be described.

  In step S902, a weak classifier to be used is determined. FIG. 10 is a diagram schematically showing the relationship between the number of weak classifiers obtained by an algorithm for boosting learning and the reliability of the corresponding weak classifiers.

  As the number of weak classifiers increases, the reliability of the corresponding weak classifiers generally decreases, but does not always decrease uniformly, but decreases on average while repeating up and down.

  Therefore, in step S902, when determining the weak classifier to be used, the weak classifier is determined based on a predetermined reliability. For example, in the case of the example shown in FIG. 10, a weak classifier having a reliability exceeding the threshold Th is used for actual processing, and a weak classifier equal to or less than Th is not used for rectangular filter calculation processing.

  By specifying the weak discriminator to be used in this way, the weak discriminator having higher reliability than the case of determining the weak discriminator to be uniformly used according to the number of processes as in the first embodiment. It is possible to specify (a classifier having a high contribution rate).

  Here, the reliability of the weak discriminator string generated by learning is determined, the weak discriminator having a predetermined condition is set as “weak discriminator to be used”, and the weak discriminator not having the predetermined condition is not used. Mask information for “weak classifier” is generated in advance. That is, the mask information here is information that describes whether or not the weak classifier is used.

FIG. 11 is a diagram illustrating an example of a table that associates the operation clock with the total number of weak classifiers to be used. In FIG. 11, N1 to N3 and Th _xy (x: clock type, y: stage number) are the same variables as in the first embodiment. M _xy (x: clock type, y: stage number) is mask information for designating whether or not each weak discriminator is executed, and is represented by a binary bit string. An example of _Mxy is shown below. Here, an example is shown in which 32 weak discriminators are assigned in ascending order from MSB (Most Significant Bit).
M11 = 11111111111111111010100101010000
The weak classifier assigned 1 in each bit is actually used during the rectangular filter calculation process, and the weak classifier assigned 0 is not used during the rectangular filter calculation process. For example, the first weak classifier is used for the rectangular filter calculation process, and the 32nd weak classifier is not used.

In step S902, a predetermined M _xy is selected according to the operation clock, and is set in a register (not shown) in the process execution determination unit 203.

  The processing shown in steps S903 to S907 is the same as the processing shown in steps S703 to S707 in FIG.

  Further, only the difference from the processing of step S708 to step S719 in FIG. 7 will be described for the processing of step S908 to step S919 as the operation of the identification processing unit 103.

In step S920, the process execution determination unit 203 determines a weak classifier to be used. Based on the rectangular filter processing number counter counted by the processing number counter 202 and the mask information M _xy set in step S902, the processing execution determination unit 203 determines whether or not to use for the rectangular filter calculation processing.

That is, the corresponding bit position of _Mxy is determined from the value of the rectangular filter processing number counter, and it is determined whether the value is 0 or 1. When it is determined that the value of the bit position corresponding to the value of the rectangular filter processing number counter is 1, it is determined that the weak classifier is a “weak classifier to be used”, and steps S909 to S911 are performed. A rectangular filter calculation process and a weak discrimination process are executed.

  In step S912, the rectangular filter processing number counter is updated regardless of whether the weak classifier is used. In step S913, it is determined whether or not the rectangular filter calculation processing for all weak discriminator rows in one stage has been completed. This determination is made based on a comparison between the number of weak classifiers generated by learning in advance and a rectangular filter processing number counter.

  The subsequent processing in steps S913 to S919 is the same as the processing in steps S513 to S519 in FIG.

  As described above, in the present embodiment, the number of weak classifiers to be used is changed in accordance with the operation clock of the pattern identification apparatus that executes the identification process. This makes it possible to adjust the extraction accuracy and extraction processing speed according to the operation clock of the embedded device even when the identification processing unit that has been implemented as hardware as a common device is mounted on multiple embedded devices with different specifications. It becomes.

  Furthermore, in the present embodiment, when the weak classifier to be used is specified, it is specified according to the reliability of the weak classifier. Thereby, compared with the said 1st Embodiment, it becomes possible to suppress degradation of extraction accuracy.

[Third Embodiment]
In the first and second embodiments, the tradeoff between the extraction accuracy and the extraction processing speed is adjusted by changing the number of weak classifiers used according to the operation clock of the pattern identification device. The present invention is not limited to this. For example, the trade-off between the extraction accuracy and the extraction processing speed may be adjusted by dynamically changing the number of weak classifiers to be used according to the extraction status of the object to be extracted.

  Hereinafter, the details of the pattern identification apparatus according to the present embodiment for adjusting the trade-off between the extraction accuracy and the extraction processing speed will be described. Note that the hardware configuration of the pattern identification apparatus according to the present embodiment, the detailed configuration of the rectangular filter calculation unit 201, the configuration of the parameter information 21, and the like are the same as those in the first embodiment, and a description thereof will be omitted here. .

  FIG. 12 is a flowchart showing the operation of the pattern identification apparatus 100 according to the third embodiment of the present invention. Hereinafter, only the differences from the second embodiment will be described for the sake of simplicity.

  The processing from step S1201 to step S1220 is the same as the processing from step S901 to step S920 in FIG. However, the processing contents of steps S1202, S1203, and S1217 are different from those of steps S902, S903, and S917 in FIG. 9, respectively.

  In step S1221, the number of extraction target objects extracted in the previous frame is read. Next, in steps S1202 and S1203, the number of weak discriminators to be used and the number of stages are determined based on the number of objects to be extracted in the previous frame read here. For example, when the number of extraction target objects extracted in the previous frame is large, the number of weak classifiers to be used is decreased.

FIG. 13 is a diagram illustrating an example of a table associated with the number of extraction target object extractions of the previous frame and the total number of weak discriminators to be used, which the pattern identification apparatus according to the present embodiment has. In FIG. 13, N1 to N3, Th _xy (x: clock type, y: stage number) and M _xy (x: clock type, y: stage number) correspond to the variables shown in FIG. 11 of the second embodiment. .

  In the case of the table shown in FIG. 13, three parameters are stored corresponding to the number of extraction target objects extracted in the previous frame. In FIG. 13, D1 and D2 correspond to the number of extracted objects to be extracted. In this embodiment, when the number of extracted extraction target objects is large, the total number of weak classifiers is set to decrease. 13 is stored in the ROM 109.

  In step S1202 and step S1203, the number of extraction target objects extracted in the previous frame (step S1221) is compared with the conditions (D1, D2) described in the table. As a result, the number of stages used is changed.

That is, when the number of objects to be extracted changes, in steps S1202 and S1203, the total number of weak discriminators and parameters (N1 to N3, Th _xy and M _xy ) corresponding thereto are registered in the processing execution determination unit 203. Set to.

  In step S1204 to step S1220, identification processing is executed according to the parameters set here.

  The processing of steps S1204 to S1220 is equivalent to the processing of steps S904 to S920 shown in the second embodiment, but in this embodiment, the number of extracted extraction target objects is counted in step S1217. Is different.

  In step S1217, the number of extracted objects to be extracted (count value after completion of frame processing) is read out in step S1221 and used in the next frame identification processing. It is assumed that the counter is initialized after being read in step S1221.

  As described above, in the present embodiment, the number of weak classifiers to be used is dynamically changed according to the extraction result in the previous frame. As a result, when the number of extraction target objects extracted in the previous frame is small, high-precision extraction is executed by many weak classifiers. Further, when the number of extraction target objects extracted in the previous frame is large, the number of weak classifiers is reduced and the extraction processing speed priority process is executed. As a result, even when it is necessary to execute the extraction process on the moving image in real time within a predetermined time, the extraction process can be appropriately realized.

[Fourth Embodiment]
In the third embodiment, the trade-off between the extraction accuracy and the extraction processing speed is adjusted by changing the number of weak classifiers used according to the extraction result in the previous frame. Not limited to. For example, the tradeoff between the extraction accuracy and the extraction processing speed may be adjusted by changing the number of weak discriminators to be used according to the extraction situation in the current frame being processed.

  Hereinafter, the configuration of the pattern identification device according to the present embodiment will be described. Note that the hardware configuration of the pattern identification apparatus according to the present embodiment, the detailed configuration of the rectangular filter calculation unit 201, the configuration of the parameter information 21, and the like are the same as those in the first embodiment, and a description thereof will be omitted here. .

  FIG. 14 is a flowchart showing the operation of the pattern identification apparatus 100 according to the fourth embodiment of the present invention. Hereinafter, only the differences from the third embodiment will be described for the sake of simplicity.

  The processing in steps S1401 to S1421 is substantially equivalent to the processing in steps S1201 to S1221 in FIG. 12, but in this embodiment, the processing order and the processing contents in step S1421 are different.

  When the SAT data is stored in the RAM 204 as processing target image data in steps S1404 to S1407, identification processing for the processing target frame is started.

  In step S1421, the number of extraction target objects extracted so far is read. The read value is a counter value that is counted up every time the extraction target object is extracted in step S1417. In step S1421, the counter is initialized when the first processing block in the frame is processed.

  In steps S1402 and S1403, the weak classifier to be used is determined according to the read count value (the number of extraction target objects extracted in the frame). The weak classifier is determined with reference to the table shown in FIG. 13 as in the third embodiment.

  That is, it is determined whether or not a change has occurred in the conditions shown in the table. If it is determined that a change has occurred, the corresponding parameters are set in steps S1402 and S1403.

  In steps S1408 to S1418, identification processing for the processing block is executed, and when an extraction target object is extracted, a counter that counts the number of extractions is updated in step S1417. When the processing of the processing block that is a partial image is completed, in step S1421, step S1402, and step S1403, the number of weak classifiers to be used is reset again based on the table.

  As described above, in the present embodiment, the number of weak classifiers is reset for each processing block serving as a partial image. That is, the number of weak classifiers is dynamically changed in units of processing blocks according to the number of extraction target objects extracted in the frame currently being processed. As a result, if the number of objects to be extracted in the frame currently being processed increases and there is a possibility that the extraction processing time may run out, the number of weak classifiers is reduced when processing subsequent processing blocks. It is possible to give priority to shortening the extraction processing time.

  As a result, even when it is necessary to execute the identification process on the moving image in real time within a predetermined time, the identification process can be appropriately executed.

[Fifth Embodiment]
In the first embodiment and the second embodiment, the tradeoff between the extraction accuracy and the extraction processing speed is adjusted by changing the number of weak classifiers to be used according to the operation clock. The present invention is not limited to this. For example, the number of weak classifiers to be processed may be changed according to the shooting mode designated by the user.

  Hereinafter, the configuration of the pattern identification device according to the present embodiment for adjusting the trade-off between the extraction accuracy and the extraction processing speed will be described. Note that the hardware configuration of the pattern identification apparatus according to the present embodiment, the detailed configuration of the rectangular filter calculation unit 201, the configuration of the parameter information 21, and the like are the same as those in the first embodiment, and a description thereof will be omitted here. .

  The operation of the pattern identification apparatus 100 will be described using the flowchart (FIG. 9) in the second embodiment.

  In general, in the case of a camera or the like, a plurality of shooting modes can be set according to an object to be shot. FIG. 15 is a diagram for explaining an example of the shooting mode, where (a) is a portrait mode and (b) is a diagram showing a typical composition for shooting in the standard mode.

  Reference numerals 1501 and 1502 schematically represent subjects. In the portrait mode, a few subjects 1501 are photographed relatively large, and the background other than the subject is blurred (for example, the optical system of the image input unit 101 is controlled to reduce the depth of field).

  In this way, it is possible to predict the composition to some extent according to the shooting mode. That is, when the subject is detected from the composition, the number of extraction target objects can be predicted. For example, in the portrait mode, since the number of subjects is small and the background area is blurred, the number of extraction target objects to be extracted is small.

  In this embodiment, the trade-off between extraction accuracy and extraction processing speed is adjusted from such a viewpoint. Specifically, when the user designates a shooting mode in the I / F unit 111, the number of weak classifiers to be actually used is determined in step S902 and step S903 according to the designated shooting mode.

  FIG. 16 is a diagram illustrating an example of a table associating a shooting mode with the total number of weak classifiers to be used, and is used when determining weak classifiers to be used in steps S902 and S903.

N1 to N2, Th _xy (x: clock type, y: stage number) and M _xy (x: clock type, y: stage number) are the same as those in the table (FIG. 11) described in the second embodiment. In this case, each bit of N1> N2 and _M1y is designated as 1 or the like.

  That is, in the portrait mode, identification is performed using more weak classifiers (all weak classifiers generated in advance by learning), and in other cases, the number of weak classifiers is reduced. Assume that the table is stored in the ROM 109. In other steps, identification processing is performed according to the conditions set here. Since other steps are the same as those in the second embodiment, description thereof is omitted.

  As described above, in this embodiment, in the portrait mode, all weak classifiers determined at the time of learning are used to perform high-precision discrimination processing, and in the standard mode, the number of weak classifiers is reduced and high-speed processing is performed. It was set as the structure which performs.

  This makes it possible to control the number of weak classifiers used in accordance with the shooting mode, and to suitably adjust the trade-off between extraction accuracy and extraction processing speed.

[Sixth Embodiment]
In the seventh embodiment, the number of weak classifiers to be used is controlled according to the shooting mode. However, the present invention is not limited to this. For example, it is good also as a structure which changes the number of weak discriminators used according to the operation mode which a user designates.

  Depending on the operation mode of the imaging device, the extraction target object is extracted in real time in parallel with the generation of the image data, and the non-real time extraction is performed from the image data that has been generated and stored in the recording unit of the imaging device. Sometimes the target object is extracted. In general, the former is called an online mode and the latter is called an offline mode. The online mode corresponds to the shooting operation, and the offline mode corresponds to the operation mode when, for example, it is desired to perform some processing / search on the image data after shooting.

  For this reason, the extraction processing speed has priority in the online mode, and the extraction accuracy has priority in the offline mode. In this embodiment, the number of weak discriminators to be used is controlled according to the operation mode of the pattern discriminating apparatus. The operation of this embodiment will also be described using the flowchart (FIG. 9) of the second embodiment.

  Specifically, when the user designates an operation mode in the I / F unit 111, the number of weak classifiers to be used is determined in accordance with the designated operation mode in steps S902 and S903.

  FIG. 17 is a diagram illustrating an example of a table that associates the operation mode with the total number of weak classifiers to be used, and is used when determining weak classifiers to be used in step S902 and step S903.

N1 to N2, Th _xy (x: clock type, y: stage number) and M _xy (x: clock type, y: stage number) are the same as those in the table (FIG. 11) described in the second embodiment.

  In FIG. 17, the parameters are set so that the discrimination processing is performed using all weak classifiers generated during learning in the offline mode, and the classification processing is performed by reducing the number of weak classifiers in the online mode. Yes. Assume that the table is stored in the ROM 109. In other steps, identification processing is performed according to the conditions set here. Since other steps are the same as those in the second embodiment, description thereof is omitted.

  As described above, in the present embodiment, in the offline mode, all weak discriminators determined at the time of learning are used to perform high-precision discrimination processing, and in the online mode, the number of weak discriminators is reduced and high-speed discrimination processing is performed. It was set as the structure which performs.

  This makes it possible to control the number of weak classifiers to be used according to the operation mode, and to suitably adjust the tradeoff between extraction accuracy and extraction processing speed.

[Seventh Embodiment]
In the first to sixth embodiments, the parameter information 211 that defines the operation of the weak discriminator is stored in the ROM 205 when adjusting the trade-off between the extraction accuracy and the extraction processing speed. However, the present invention is not limited to this.

  In the present embodiment, a case will be described in which parameter information that defines the operation of the weak classifier is stored in a mixed recording unit (for example, ROM) and a changeable recording unit (for example, RAM).

  FIG. 18 is a diagram schematically illustrating parameter information that defines the operation of the weak classifier stored in the ROM 205 and the RAM 204 and the usage status thereof. FIG. 18A shows a state where the parameter information of the weak classifier generated by learning is stored in the ROM 205.

  One box is a data set of parameter information that defines one rectangular filter stored in a predetermined area on the ROM 205. That is, it is assumed that parameter information regarding one rectangular filter shown in FIG. 4 is stored in one box. In addition, parameter information regarding a plurality of rectangular filters to be processed in order is stored in the order of boxes from the left.

  FIG. 18B shows the actual operation of the weak classifier described in the first embodiment. This indicates that the rectangular filter corresponding to the shaded box is a rectangular filter used for actual processing. In this case, the number of weak classifiers (rectangular filters) is controlled by ending the rectangular filter row generated by learning at an early stage.

  FIG. 18C shows the actual operation of the weak classifier described in the second embodiment. In this case, the number of weak classifiers (rectangular filters) actually used is controlled by performing processing so as to skip a specific rectangular filter having low reliability.

  This embodiment is characterized in that it operates as shown in FIGS. 18 (d) and 18 (e). In other words, the parameter information constituting the rectangular filter is held not only in the ROM 205 (first storage means) but also in the RAM 204 (second storage means). Further, after the predetermined position (hereinafter referred to as a “branch point”) or for a predetermined section, the rectangular filter arithmetic processing is executed using the parameter information stored in the RAM 204.

  FIG. 19 is a diagram showing an example of a memory map of the ROM 205 and RAM 204 viewed from the rectangular filter calculation unit 201. Parameter information created by learning in advance is stored in the range of ADR0 to (ADR1-1).

  On the other hand, addresses above ADR1 are secured in the RAM 204, so that additional parameter information for performance improvement can be stored after shipment of the pattern identification device. When the parameter information to be referred to is changed from the ROM 205 to the RAM 204 at the branch point, an address value (= value after ADR1) of the RAM 204 area is set in a pointer register described later.

  Hereinafter, the details of this embodiment will be described with reference to the flowchart shown in FIG. For simplification of description, only the difference from the first embodiment will be described here.

  In step S2002, a condition for designating a rectangular filter row to be processed is set based on the operation clock. Specifically, the number of rectangular filters where the branch occurs (equivalent to the rectangular filter number) and the address for storing the corresponding parameter information are set in the register in the process execution determination unit 203. In the following, a point at which loading of parameter information starts from a physically different storage destination is referred to as a branch.

  FIG. 21 is a diagram illustrating a part of the processing execution determination unit 203 unique to the present embodiment. Reference numerals 2111 to 211n denote pointer addresses, which are registers that store head addresses that refer to parameter information that defines the rectangular filter. 2121 to 212n are registers for designating branch points of the rectangular filter row, and correspond to the rectangular filter numbers.

  Reference numeral 2130 denotes a comparator which compares the branch point register with the processed register counter to determine the branch point. If it is determined that the branch point has been reached, an address load signal and an address are output.

  Reference numeral 2140 denotes a selector that selects a branch destination address corresponding to the branch point from pointer addresses 2111 to 211n according to the output of the comparator. In step S2002, branch point registers 2121 to 212n and pointer addresses 2111 to 211n are set.

  In the case of the example shown in FIG. 18D or 18E, one branch point register and one pointer address are set for each stage.

FIG. 22 is a diagram showing an example of a table that associates the operation clock and the total number of weak classifiers to be used in this embodiment. In FIG. 22, N1 to N3 and Th _xy (x: clock type, y: stage number) are the same variables as in the first embodiment.

AD _xy (x: clock type, y: stage number) is pointer information for designating a branch destination address, and P _xy (x: clock type, y: stage number) is a rectangular filter number for branching.

In step S2002, the table is referenced according to the operation clock, AD _xy is set to pointer addresses 2111 to 211n, and P _xy is set to branch point registers 2121 to 212n.

  In the case of the example shown in FIG. 18D, the parameter information of the rectangular filter is further added to the parameter information of the rectangular filter that is generated by learning in advance and hardened as the ROM 205. In step S2021, parameter information of an additional rectangular filter for improving the extraction accuracy is loaded into the RAM 204, which is the branch destination address specified in step S2002.

  The parameter information of the additional rectangular filter is data stored in the ROM 205 or the like, and is transferred to the RAM 204 by the CPU 108 in this step.

  When the identification processing unit 103 is operated with a high-speed operation clock, there is a margin in processing time. Thus, the number of rectangular filters is increased in this way to improve extraction accuracy.

  In the case of the example shown in FIG. 18E, a part of the parameter information of the rectangular filter generated by learning in advance and hardened as a ROM is replaced. In step S2021, parameter information of the exchange rectangular filter for increasing the extraction accuracy or the extraction processing speed is loaded into the RAM 204 that is the branch destination address designated in step S2002. The parameter information is data stored in the ROM 109 or the like, and is transferred to the RAM 204 by the CPU 108 in this step.

  With such a configuration, even after the identification processing unit 103 is implemented as hardware and the contents of the ROM 205 are fixedly determined, part of the parameter information of the rectangular filter can be replaced. The parameter information to be replaced here may be parameter information newly learned for improving performance, or may be parameter information newly learned by adding the extraction processing speed as a constraint condition. In this way, it is possible to adjust the tradeoff between extraction accuracy and extraction processing speed time by partially replacing the parameter information.

  In step S2003, the number of stages is determined based on the operation clock. The process here is the same as step S503.

  Next, in step S2004 to step S2007, the image data to be processed is transferred to the RAM 204. In step S2008, a processing block is determined, and in step S2009, reading of parameter information is started.

  A cumulative value is calculated based on an evaluation value that is a result of a predetermined rectangular filter calculation process (steps S2010 to S2011). In step S2012, the number of processed rectangular filters is updated.

  In step S2020, the branch of the rectangular filter row is determined. Specifically, the determination is made by comparing the filter processing number counter with the designated branch point filter number by the comparator 2130 shown in FIG.

  If it is determined that it is a branch point, in step S2021, the parameter information reference destination head address is newly set. Specifically, the pointer address 2111 shown in FIG. 21 is loaded into a memory access address counter (not shown) of the rectangular filter arithmetic unit 201 according to the address load signal.

  Thereafter, parameter information necessary for the rectangular filter calculation processing is processed with reference to the area set here. Here, it is assumed that the head address of the RAM 204 is stored in the pointer address 2111 (step S2002).

  The rectangular filter calculation unit 201 continues the processing while sequentially reading out parameter information that defines the rectangular filter with the reset address as the head address (steps S2009 to S2013).

  Note that the type shown in FIGS. 18D and 18E is the relationship between the number of weak classifiers generated in advance and stored in the ROM 205 and the number of weak classifiers specified for use in extraction. The structure is the same.

  Hereinafter, step S2013 to step S2019 are equivalent to step S513 to S519 of the first embodiment, and thus description thereof is omitted.

  As described above, in the present embodiment, it is possible to flexibly replace the weak discriminator series in accordance with a predetermined precondition.

  In the present embodiment, the case of loading parameter information to be added and changed in step S2021 has been described. However, the parameter information may be loaded for each stage. In that case, parameter information is loaded into the RAM 204 when it is determined that one stage has ended (step S2013). In this case, when there are a plurality of stages, the RAM 204 is shared for each stage. Therefore, the present embodiment can be realized with a smaller number of RAMs 204.

  Moreover, although the case where there is one branch point has been described in the examples illustrated in FIGS. 18D and 18E, the number of branch points is not limited to one, and a plurality of branch points may be provided. In this embodiment, the case where the weak discriminator string is branched in units of weak classifiers in the stage has been described. However, the present invention is not limited to this, and it may be configured to branch in units of stages. That is, the storage destination of all parameter information necessary for all rectangular filter calculation processes in a specific stage may be stored in the RAM 204.

  As described above, according to the present embodiment, it is possible to improve performance and adjust trade-offs even after the identification processing unit 103 is created as a hardware device. Further, it can be realized on a smaller scale than a configuration in which all parameter information is stored as a RAM.

[Eighth Embodiment]
In the fifth and sixth embodiments, the configuration in which the number of weak classifiers to be used is determined in conjunction with the shooting mode and the operation mode has been described. However, the present invention is not limited to this, and the user directly sets the extraction condition. It is good also as a structure to select. In that case, the user designates the extraction mode via the I / F unit 111.

FIG. 23 shows an example of a table for determining extraction conditions when there are two extraction modes. N1 to N2, Th _xy (x: clock type, y: stage number) and M _xy (x: clock type, y: stage number) are the same as those in the second embodiment. By creating such a table in advance, the user can directly adjust the trade-off between extraction accuracy and extraction processing speed.

  Moreover, although each said embodiment demonstrated the case where the method currently disclosed by the nonpatent literature 1 was applied to this invention, this invention is not limited to this. The present invention can be applied to various identification processes for performing identification based on a plurality of weak classifiers connected in cascade.

  Further, although cases have been described with the above embodiments where a rectangular filter is used as the weak classifier, the present invention is not limited to this, and other various weak classifiers can be applied.

  Further, although cases have been described with the above embodiments where the number of stages and the number of rectangular filters in the stage are changed, the present invention is not limited to this. The number of stages may be fixed and only the number of weak classifiers in the stage may be changed. Also, depending on the algorithm, there may be one stage, but the present invention is also applicable to such a system.

  In each of the above embodiments, the case where a specific object is extracted from image data has been described. However, the present invention is not limited to this, and the present invention is applied to an apparatus that identifies a specific object (pattern) from a one-dimensional signal such as an audio signal. It is also possible to do.

In the second embodiment, the mask information for preliminarily specifying the weak classifier to be used is created based on the reliability. However, the present invention is not limited to this. For example, instead of setting the mask information M _xy shown in FIG. 11 in the process execution determination unit 203, a reliability threshold may be directly set in the process execution determination unit 203. Furthermore, it is good also as a structure which determines whether the reliability of parameter information is compared with the reliability set in step S902 each time, and a corresponding rectangular filter is used.

  In each of the above embodiments, the case where the identification processing unit 103 is realized by a hardware device has been described. However, a configuration in which the identification processing unit 103 is realized by software such as a DSP (Digital Signal Processor) may be used.

[Other Embodiments]
Note that the present invention can be applied to a system (for example, a copier, a facsimile machine, etc.) consisting of a single device even when applied to a system composed of a plurality of devices (for example, a host computer, interface device, reader, printer, etc.). You may apply.

  Needless to say, the object of the present invention can also be achieved by supplying a system or apparatus with a recording medium that records a program code of software that implements the functions of the above-described embodiments. In this case, the function is realized by the computer (or CPU or MPU) of the system or apparatus reading and executing the program code stored in the recording medium. In this case, the recording medium on which the program code is recorded constitutes the present invention.

  As a recording medium for supplying the program code, for example, a floppy (registered trademark) disk, hard disk, optical disk, magneto-optical disk, CD-ROM, CD-R, magnetic tape, nonvolatile memory card, ROM, or the like is used. be able to.

  Further, the present invention is not limited to the case where the functions of the above-described embodiments are realized by executing the program code read by the computer. For example, an OS (operating system) running on a computer performs part or all of actual processing based on an instruction of the program code, and the functions of the above-described embodiments may be realized by the processing. Needless to say, it is included.

  Further, when the program code read from the recording medium is written in a memory provided in a function expansion board inserted into the computer or a function expansion unit connected to the computer, the functions of the above-described embodiments are realized. Is also included. That is, after the program code is written in the memory, the CPU or the like provided in the function expansion board or function expansion unit performs part or all of the actual processing based on the instruction of the program code, and is realized by the processing. This is also included.

It is a figure which shows the structural example of the information processing apparatus (pattern identification device 100) concerning the 1st Embodiment of this invention. 5 is a diagram for explaining details of an identification processing unit 103. FIG. 3 is a diagram illustrating a detailed configuration of a rectangular filter calculation unit 201. FIG. It is a figure which shows an example of the parameter information 211. It is a flowchart which shows operation | movement of the pattern identification apparatus 100 concerning the 1st Embodiment of this invention. It is a figure which shows typically the relationship between the total number of weak discriminators (rectangular filter) and extraction accuracy. It is a figure which shows an example of the table which associates an operation clock and the total number of weak discriminators to be used. It is a figure for demonstrating the method to calculate a rectangular block from SAT data. It is a flowchart which shows operation | movement of the pattern identification apparatus 100 concerning the 2nd Embodiment of this invention. It is the figure which showed typically the relationship between the number of weak classifiers obtained by an algorithm for boosting learning, and the reliability of a corresponding weak classifier. It is a figure which shows an example of the table which associates an operation clock and the total number of weak discriminators to be used. It is a flowchart which shows operation | movement of the pattern identification apparatus 100 concerning the 3rd Embodiment of this invention. It is a figure which shows an example of the table which correlates the extraction object extraction number of the previous frame, and the total number of weak discriminators to be used. It is a flowchart which shows operation | movement of the pattern identification apparatus 100 concerning the 4th Embodiment of this invention. It is a figure for demonstrating imaging | photography mode. It is a figure which shows an example of the table which associates imaging | photography mode and the total number of weak discriminators to be used. It is a figure which shows an example of the table which associates an operation mode and the total number of weak discriminators to be used. It is a figure which illustrates typically the parameter information which prescribes | regulates the operation | movement of the weak discriminator stored in ROM205 and RAM204, and its use condition. 3 is a diagram illustrating an example of a memory map of a ROM 205 and a RAM 204. FIG. It is a flowchart which shows operation | movement of the pattern identification apparatus 100 concerning the 7th Embodiment of this invention. It is a figure for demonstrating a part of process execution determination part 203 of the pattern identification device 100 concerning the 7th Embodiment of this invention. It is a figure which shows an example of the table which associates an operation clock and the total number of weak discriminators to be used. It is a figure which shows an example of the table for extraction condition determination in the case of having two extraction modes. It is a figure for demonstrating the method of connecting a discriminator which consists of a some weak discriminator group produced | generated by the boosting learning algorithm to a cascade, and advancing a process, performing an abort determination for every discriminator. It is a figure which shows the example of the rectangular filter for extracting a specific object. It is a figure showing the image data of a process target.

Claims (11)

A plurality of weak classifiers connected to the cascade to process the data,
First storage means for storing information defining the processing content of each of the plurality of weak classifiers;
Preconditions for processing the data and information for determining weak classifiers to be used for processing the data among the plurality of weak classifiers are stored in a table defined for each precondition. Second storage means for
Selecting means for selecting, from the plurality of weak classifiers, weak classifiers to be used for processing the data by referring to the table;
Based on the information that defines the processing content corresponding to the weak classifier selected by the selection means, an evaluation value obtained by processing the data by the weak classifier is used. An information processing apparatus comprising: extraction means for extracting In the table,
The number of weak classifiers to be used is defined in units of stages formed by connecting a plurality of weak classifiers in cascade according to the processing speed of the extraction unit. Information processing device. In the table,
2. The information for designating weak classifiers to be used according to the processing speed of the extraction means is defined for each stage formed by connecting a plurality of weak classifiers in cascade. The information processing apparatus described.   4. The information processing according to claim 3, wherein the information specifying the weak classifier to be used is generated based on a reliability of the weak classifier included in the information defining the processing content. apparatus. In the case where the data is image data composed of a plurality of frames,
In the table,
The number of weak classifiers to be used is defined in units of stages formed by connecting a plurality of weak classifiers in cascade according to the number of objects extracted from the image data of the previous frame by the extraction means. The information processing apparatus according to claim 1. In the case where the data is image data generated by an imaging device,
In the table,
Information specifying the weak discriminator to be used is specified for each stage formed by connecting a plurality of weak discriminators in cascade according to the shooting mode when the imaging device generates image data. The information processing apparatus according to claim 1. In the case where the data is image data generated by an imaging device,
In the table,
Whether it is a mode for processing image data in parallel with an operation for generating image data by the imaging device, or a mode for processing image data after the operation for generating image data by the imaging device is completed Accordingly, the number of weak classifiers to be used is defined in units of stages formed by connecting a plurality of weak classifiers in cascade. The table defines the number of stages to be used,
The extraction means extracts an object from the data by using the accumulated value of the evaluation values of the weak classifiers obtained for each stage by processing the data under the specified number of stages. The information processing apparatus according to claim 2, wherein the information processing apparatus is an information processing apparatus. The first storage means includes a non-rewritable first storage means and a rewritable second storage means,
The extraction unit extracts an object from the data based on information defining the processing content stored in either the first storage unit or the second storage unit according to the precondition. The information processing apparatus according to any one of claims 1 to 8. An information processing method in an information processing apparatus that processes data using a plurality of weak classifiers connected in cascade,
A first storage step of storing information defining the processing content of each of the plurality of weak classifiers;
Preconditions for processing the data and information for determining weak classifiers to be used for processing the data among the plurality of weak classifiers are stored in a table defined for each precondition. A second storing step,
By selecting the weak discriminator to be used for processing the data from the plurality of weak discriminators by referring to the table,
Using the evaluation value of the weak classifier obtained by processing the data based on the information defining the processing content corresponding to the weak classifier selected in the selection step, the object from the data An information processing method comprising: an extraction step for extracting. A program for causing a computer to execute the information processing method according to claim 10.

Images